VigiDoc a product by SSHack

Guide

ChatGPT and consumer AI at work: what a DLP sees, and what it doesn't

A salesperson pastes a draft contract into ChatGPT to get a summary. A developer submits a function that keeps crashing. An assistant uploads a customer spreadsheet to remove duplicates. None of them thinks they are taking data out. This guide brings together the published figures, OpenAI's terms on reusing conversations, what a detection-based DLP (data loss prevention) tool sees of these uses and what it misses, and then a five-step method for an SMB. It is written by the team at SSHack, an offensive security firm and the publisher of VigiDoc.

By the SSHack team · Published · about 15 minutes to read

Key takeaways

According to Verizon's 2026 DBIR, which draws on its contributors' DLP tools, 45% of employees now use AI regularly on their work devices, up from 15% a year earlier, and 67% of users go through a non-corporate account. In France, 91% of the chief information security officers (CISOs) surveyed by Proofpoint fear customer data loss through public generative AI tools. An endpoint DLP that reads browser history sees the visit to chatgpt.com or claude.ai, with the time and the device, and can raise an alert on the name of the service; it does not see the pasted text, the uploaded file or the answer. That trace helps you measure actual use, choose the tool to provide and write a rule that employees can follow.

The figures

Three sources, three angles: events detected by DLP tools (Verizon), statements from CISOs (Proofpoint), and statements from a regulated profession (the CNB, France's national bar council). Each one is given with its edition and scope.

45%

of employees use AI regularly on their work devices, whether authorized or not; a year earlier, the figure was 15%.

Verizon, Data Breach Investigations Report 2026 · contributors' DLP data, 2025, worldwide

67%

of users access AI services with a non-corporate account from their work device; 72% the previous year.

Verizon, DBIR 2026 · same scope

91%

of French CISOs surveyed are concerned about customer data loss through public generative AI tools; 92% think employees may use AI in ways that expose sensitive data.

Proofpoint, Voice of the CISO 2026 · 100 French CISOs, organizations with 1,000+ employees, May 2026

1 in 10

lawyers trusts generative AI to guarantee data protection; among firms that use it, 23% rely only on general-purpose tools.

CNB, Les cahiers de l'Observatoire, L'intelligence artificielle, 2025 · 4,457 lawyers surveyed

Verizon's figure must be read with its definition: a regular user is someone who opens an AI platform at least once every fifteen days, and the measurements come from contributors' DLP tools, so from companies that are already equipped. The sample does not represent French SMBs. The same report ranks unauthorized AI use third among the non-malicious insider actions detected in 2025, with a share four times higher than a year earlier. It also counts more than 15% of users with an unauthorized AI browser extension in the average organization.

Proofpoint's figure comes from a vendor survey: it measures stated concern, not observed leaks. In the same survey, 95% of French CISOs say they block or restrict the use of generative AI tools, and 90% consider their controls effective. Those answers sit alongside the 67% of personal accounts measured by Verizon. For broader statistics, see our 2026 statistics page, generative AI section (in French).

The uses that take data out

Based on the DLP events analyzed by Verizon, source code is by far the most common type of data sent to external AI tools, ahead of images and other structured data. In 3.2% of these events, it was technical and research documentation. Translated into everyday office tasks, the list is familiar.

None of these actions is malicious, which is exactly why they are so common. Employees use the most convenient tool available to them, often because the company does not provide one. In its 2025 cyber threat overview, ANSSI (France's national cybersecurity agency) notes that integrating generative AI services into workflows is likely to increase the attack surface "in the absence of physical or usage-level separation" (our translation).

Samsung, 2023: a case reported by the press

In May 2023, Samsung banned generative AI on company computers, tablets and phones and on its internal networks, after discovering that employees had uploaded internal source code to ChatGPT. The information comes from an internal memo reported by Bloomberg and picked up by the French press. The exact nature of the code was not made public. The case illustrates the top use measured by Verizon, and a common reaction among large companies: ban first.

Where the text goes: personal or business account

Once pasted, the text is processed on the service provider's servers. What happens next depends on the type of account, and this is where the 67% of non-corporate accounts becomes meaningful. We read the pages of OpenAI, the company behind ChatGPT, on October 21, 2026; other services have their own terms, to be read the same way.

Reuse of conversations for model training by ChatGPT account type, based on OpenAI's pages accessed October 21, 2026
Account typeModel trainingWhat the company controls
Individual plan (Free, Go, Plus, Pro)Possible: the European privacy policy states that OpenAI may use the content provided to train its models, unless the user turns off the "Improve the model for everyone" settingNothing: the account belongs to the employee, who keeps it, with its full history, after leaving
Temporary chat, on an individual planNo, according to OpenAI; the conversation may be kept for up to 30 days for safety reasonsNothing: it is the user's choice, one conversation at a time
Use without an accountPossible; according to OpenAI, the training setting can be turned off in Settings, and the choice is saved in that browserNothing
ChatGPT Business, Enterprise or EduNot by default: OpenAI states that it does not use the content of these workspaces to train its modelsThe contract, the workspace, the user accounts and, depending on the plan, the retention period

Two practical consequences. Banning "ChatGPT" without distinguishing between accounts makes little sense, because the main risk lies in the personal account; providing a business account covered by a contract changes the situation more. And the training setting of a personal account depends on each employee, whose history follows them when they leave the company: a contract summarized on a Tuesday is still readable in their account months later.

French authorities give the same advice. Generative AI fact sheet no. 4, published by France Num (the French government's digital program for small businesses) and the CNIL for very small businesses and SMBs (November 2024), states that anything entered into consumer online tools may be used as training data. It asks users not to enter confidential, personal or strategic data into them. In its first answers on deploying generative AI (July 18, 2024), the CNIL recommends favoring local, secure and specialized systems and, failing that, checking whether the provider can reuse the data entrusted to it. Whether a pasted text constitutes a personal data breach under the GDPR is decided case by case, with your data protection officer. For professional secrecy, see our page on law firms (in French).

What built-in tools see

Without a dedicated tool, an SMB sees next to nothing of this use. Windows keeps no usable log of text pasted into a web page, and its event logs do not record site visits. Browser history stays on each device, nobody looks at it, and the user can delete it. The DLP included in Microsoft 365 Business Premium covers Exchange, SharePoint and OneDrive, not workstations (see DLP and Microsoft 365 Business Premium). A firewall or proxy can record the domains contacted, if someone reads those logs. In every case, the pasted text remains invisible.

What an endpoint DLP sees

A detection-based DLP tool such as VigiDoc collects the Chrome, Edge and Firefox browser history on each Windows device, with no extension to install: for each page visited, the URL, the title and the time, linked to the device. Applied to consumer AI, that gives modest but reliable information.

VigiDoc Dashboard Settings window, Keyword alerts section: the field of comma-separated keywords (confidentiel, orion, tarif, remise, prix de revient, prix_de_revient, export crm, export_crm, sample French business terms) and their preview as tags.
Keyword settings in the VigiDoc Dashboard: a comma-separated list of terms, searched for in the names and contents of files that left the device, email subjects, and the URLs and titles of pages visited. AI service names are added like any other term. Sample data.

What it doesn't see

This list is longer than the previous one. Better to know it before buying a tool.

Two mistakes to avoid. A visit trace establishes that a service was opened; it says nothing about what was entrusted to it. And a content-inspection tool does not cover everything: DLP tools that analyze text typed into the browser require an extension or a proxy that decrypts traffic, and they do not see the phone either. The choice between these approaches is discussed in our guide DLP software: definition, how it works and how to choose.

Governing use without banning blindly: five steps

The CNIL sets out the framework in its July 2024 answers: govern use with a list of permitted and prohibited uses, train users, and involve the data protection officer, IT, the CISO and business teams from the start. The method below applies it to an SMB.

  1. Measure for a month. After informing employees (see below), add the names of AI services to your keywords and let it run. You get the services in use, the number of devices and the teams involved. The goal is to identify a need, not to find a culprit.
  2. Classify the data. Write down on one page what must never go to an external service: source code, health data, HR data, customer files, cost prices, classified documents, material covered by professional secrecy. The DGSI (France's domestic intelligence service) more broadly recommends classifying data according to its level of sensitivity; that comes before any rule on AI.
  3. Provide a tool. Heavy use reveals a real need. A business account whose contract excludes training, a service hosted in Europe or a model installed on your own servers: the choice depends on your data. An employee who has an authorized tool has fewer reasons to open their personal account.
  4. Write the rule and train people. One page appended to the IT acceptable use policy: the authorized tools, the data that is off-limits whatever the tool, the instruction to pseudonymize before pasting, what to do after a mistake. The France Num fact sheet calls for a document describing what employees can do with the system, with sample prompts. One hour of training per team, with examples from their own work, is worth more than a memo.
  5. Review every month, then restrict what remains. Are visits to unauthorized services going down now that a tool is available? Who keeps using them, and why? For services that remain problematic, filtering at the proxy or firewall is a legitimate measure in its own right; it is decided once use has been measured, on a precise scope. Our view on choosing between these two approaches is in Detect rather than block.

The order matters. A ban announced before measuring often hits the wrong service and moves usage to personal phones, where nothing is visible anymore.

Bad practices, and what they expose

The table below can serve as the basis for the policy page in step 4. The last column shows what a detection-based DLP reveals when the practice happens on an equipped Windows device.

Common uses of consumer AI at work, what they expose, and what a detection-based DLP shows
PracticeWhy it happensWhat it exposesWhat VigiDoc shows
Source code pasted into a personal account to fix an errorReal time savings, no assistant providedCode, keys and connection strings at a third party, in a history the developer keeps after leavingThe visit to the service (URL, title, time), flagged if its name is a keyword; not the pasted code
Customer contract or letter summarized by a free AI toolLong documents, tight deadlinesNames, amounts, clauses; possible reuse for training depending on the account's termsThe visit to the service; not the text
Customer export or HR file uploaded to be cleaned upSpreadsheet skills gap, urgent requestPersonal data at a third party with no contract; whether it is a data breach is a question for the DPOThe visit to the service; not the uploaded file. If the file was first copied to a USB drive or sent through Outlook, those transfers are logged
Blanket ban with no tool providedReaction to an article or an incidentUsage moves to phones or private browsing, so it becomes invisibleFewer visits in the history, which does not prove less use

Seen by an offensive team

During an internal penetration test, once it has taken control of a device, the SSHack team first looks at what the browser keeps open: email sessions, online storage, management tools and now AI services. An AI session left logged in gives access to the account's full conversation history, meaning everything its owner has pasted into it for months: contract excerpts, pieces of code, spreadsheet exports, screenshots. For an attacker, it is a summary of the company written by its own employees, sorted by date. On a personal account, the company has no way to revoke access when the employee leaves, purge the history or enforce two-factor authentication.

The second reflex is to look for what the company monitors. An AI service blocked at the proxy tells us that the rule exists and points us to the next service, the one nobody thought to block. Discreet detection, by contrast, keeps us on the path the company is watching. That is why we recommend detecting first, then blocking only the critical cases. Here, the critical cases are source code and research documents. They call for a provided tool and a written rule; a list of sites to block will always be one service behind.

What VigiDoc sees, and what it doesn't

VigiDoc, SSHack's detection-based DLP software for Windows endpoints, collects Chrome, Edge and Firefox history on each device, with no extension to install: the URL, title and time of every page visited, for all user accounts on the device. Add the names of AI services to your keywords, and every visit to chatgpt.com, claude.ai or gemini.google.com is flagged in the Dashboard, device by device. The same agent logs copies to USB drives and emails sent from Outlook, with an encrypted copy of the files on a server installed on your premises. Our common baseline of keywords is a good starting point. It does not block anything: you measure actual use before deciding.

Its limits are precise. It does not see text pasted into an AI tool or a webmail, the file uploaded to an AI tool, or the answer: only the page visit is logged. Desktop apps, browser extensions, private browsing, phones, printing, screenshots and the contents of password-protected archives are also out of its sight. The 30-day trial starts as soon as the server is installed, with no license to request, and covers the measurement month in step 1.

Inform employees, stay proportionate

Recording visits to AI services means collecting information about employee activity. In France, article L1222-4 of the French Labor Code requires employees to be informed beforehand of any system that collects information about them. The CNIL (France's data protection authority) also requires the system to be proportionate to its purpose and, in companies with 50 or more employees, the works council (CSE) must be consulted. Other countries have their own rules: check your local employment and data protection law. In practice, the IT acceptable use policy states that browsing history on company devices is recorded and that alerts cover certain services; it does not list the keywords.

Proportionality also applies to how the trace is used. A single visit justifies no disciplinary meeting; heavy use of a personal account by a team that handles source code or health data justifies a conversation, then a tool. For an individual situation that turns into a dispute, consult a lawyer: this page describes facts and tools, not a legal position.

Frequently asked questions

Is pasting text into ChatGPT a data leak?

At the very least, it is data leaving the company: the text leaves the device and the organization to be processed, and sometimes stored, on a third party's servers, depending on the service's terms. The generative AI fact sheet for small businesses published by France Num and the CNIL (France's data protection authority) states that whatever is entered into consumer online tools may be used for training. Confidential, personal or strategic data should therefore not be entered into them. If the text contains personal data, whether it counts as a breach under the GDPR is a question to settle with your data protection officer.

Are conversations from a personal ChatGPT account used to train the models?

According to OpenAI's help center, accessed October 21, 2026, conversations on individual plans (Free, Go, Plus and Pro) can be used to train the models unless the user has turned off the "Improve the model for everyone" setting. Temporary chats are not used for training, but may be kept for up to 30 days for safety reasons. For ChatGPT Business, Enterprise and Edu workspaces, OpenAI states that it does not use their content to train its models by default. These terms change: read them again on the date you write your policy.

What does DLP software see when an employee uses ChatGPT?

An endpoint DLP that reads browser history, such as VigiDoc, records every visit to an AI service in Chrome, Edge or Firefox: the URL, the page title and the time, device by device. It raises an alert if a keyword appears in the URL or the title, such as the name of the service. It does not read the text pasted into the conversation, the uploaded file or the answer. The trace tells you which devices use which AI tool and how often; it does not prove what was shared.

Should you ban ChatGPT in your company?

Not before you have measured actual use. A ban decided without measurement often targets the wrong tool and moves usage to personal phones, where the company sees nothing at all. The CNIL recommends defining a list of permitted and prohibited uses, training users and favoring local, secure and specialized systems. An SMB can measure for a month, provide a tool covered by a contract, write its rules, then restrict only what remains problematic.

Which keywords should you add to spot consumer AI use?

The names of the services, which appear in the URL or title of the pages: chatgpt, openai, claude, gemini, copilot, mistral, perplexity, deepseek. These words rarely appear elsewhere, but three cases produce irrelevant alerts: "copilot" also flags Microsoft 365 Copilot, often provided by the company, "claude" is also a first name, and "mistral" is the name of a wind. Add the services you spot during the measurement phase, and remove the ones you provide yourself.

Sources